Performances
RD Popularity Index
RD Growth Index
Comparison
Follow Updates
Follow Lists
People
About Gradient
Gradient is an open-source AI infrastructure project focused on making distributed reinforcement learning (RL) affordable. Its core stack includes Lattica, a peer-to-peer data communication protocol for connecting distributed nodes; Parallax, a distributed inference engine that turns heterogeneous devices into a unified cluster; and Echo, a distributed RL framework that decouples rollout generation from model training. Echo-2, the latest iteration, splits the RL loop into two independent swarms: a large fleet of consumer-grade devices handles the compute-light rollout generation, while a smaller cluster of datacenter GPUs (A100/H100) performs the dense training updates. Updates are relayed through a chunked peer-to-peer network to avoid bandwidth bottlenecks, and a "staleness budget" parameter lets users trade data freshness for throughput. In benchmarks, Echo-2 reduced post-training costs for a 30B-parameter model by 90%+ compared to centralized baselines, while matching or slightly exceeding learning quality. A companion protocol, VeriLLM, adds publicly verifiable inference with only ~1% overhead, addressing trust issues in permissionless networks.
The market pain point is clear: RL post-training is the most expensive step in modern AI, dominated by a few cloud providers charging premium rates for underutilized hardware. Most teams cannot afford RL fine-tuning for domain-specific models, forcing them to rely on prompt engineering with general-purpose models. Gradient's approach cuts costs by 33-90%, potentially opening RL to thousands of vertical AI teams. The project also tackles the inefficiency of idle expensive GPUs waiting for rollouts, and the difficulty of coordinating heterogeneous devices across the internet.
In the past six months (March-August 2026), Gradient released Echo-2 with peer-reviewed results, open-sourced Parallax for self-hosted distributed inference, and reported that Lattica now runs across millions of peers worldwide. The team also announced Logits, an RL-as-a-Service platform built on Echo-2, currently in development. However, the project remains pre-product-market-fit, with the stack still being proven rather than fully commoditized. The near-term user base is limited to open-source model builders, RL researchers, and infrastructure engineers who can tolerate the complexity of heterogeneous, trustless systems. No major security incidents or protocol failures were reported in this period.
Gradient Network, an open AI infrastructure layer, has marked several milestones with long-term significance. It secured a $10 million seed round led by Pantera Capital and Multicoin Capital, validating its vision of decentralized AI. A subsequent funding round with Multicoin participation further strengthened its position. The project also gained notable traction, becoming the most-followed new project on RootData within 7 days, reflecting growing community interest. These events underscore its progress toward building a peer-to-peer compute and data layer for AI, moving beyond centralized models.
Gradient (gradient.network) has made significant strides in H1 2026. Core progress includes the release of Echo-2, a distributed RL framework that decouples training from inference via a dual-swarm architecture, cutting post-training costs by 10.6x (from $4,490 to $425 on a Qwen3-30B model) while maintaining algorithmic fidelity. The team also launched Logits, an RL-as-a-Service platform that abstracts distributed orchestration, and expanded Parallax, its open-source inference engine, now adopted by leading open-model developers. Future roadmap focuses on scaling the Open Intelligence Stack (OIS) to support larger model families, enhancing heterogeneous device compatibility (including Apple Silicon and consumer GPUs), and transitioning Logits from research to production with enterprise-grade SLAs. The network is also deepening its DePIN integration, rewarding node operators for contributing idle compute, and targeting 100% hardware utilization across its global mesh.
Eric Yang (Co-founder) - Founder of Gradient. Previously Venture Investor at Sequoia Capital China (HSG) and Founding Engineer at DLive (acquired by BitTorrent). Studied Computer Science at UC Berkeley.
Yuan Gao (Co-founder) - Co-founder of Gradient. Former Head of Growth at Helium Foundation, Head of Marketing at NEO, and Managing Director at Helium China. Holds a Master's in International Finance from Columbia University and a Bachelor's in Geographic Information Science from Nanjing University.
Gradient's recent team change: Alex Mirran, Head of Business Development - North America, departed in July 2026. He joined in September 2025 and was responsible for North American market expansion and strategic partnerships. During his tenure, he drove the commercialization of products like Commonstack and participated in ecosystem collaborations. His departure may impact Gradient's business development in the North American market.